Raşid Bakır, MuhammedAtalay Çetin, MüminBakırtaş, İbrahim2025-07-102025-07-10202522148450https://dx.doi.org/10.1016/j.bir.2025.02.006https://hdl.handle.net/20.500.12451/13238This study revisits the macroeconomic determinants of non-performing loans using a deep neural network (DNN). We present the proposed DNN as a methodological framework that combines deep learning techniques and causal inference methods. We employ a rigorous triple-validation methodology that integrates deep learning architecture, random forest analysis, and the DoWhy causal inference framework. Furthermore, our optimized deep learning framework is validated and enhanced with causal inference capabilities, thus establishing a robust analytical framework for credit risk assessment in emerging markets. Although traditional analyses emphasize unemployment and debt stock as primary predictors, our causal inference methodology indicates that foreign direct investment exhibits the most substantial risk-mitigating effect. Real interest rates had substantial risk-mitigating effects compared with policy rates, suggesting the potential limitations of real interest rates in current monetary policy transmission mechanisms. The integration of deep learning and causal inference has significant implications for policy formulation, suggesting the efficacy of structural reforms over conventional monetary interventi.eninfo:eu-repo/semantics/openAccessCausal InferenceDeep LearningNon-performing LoansArticle25354155110.1016/j.bir.2025.02.006105003213293WOS:001479708900001Q1